Incremental Aggregation of Latent Semantics Using a Graph-Based Energy Model

Abstract

A graph-theoretic model for incrementally detecting latent associations among terms in a document corpus is presented. The algorithm is based on an energy model that quantifies similarity in context between pairs of terms. Latent associations that are established in turn contribute to the energy of their respective contexts. The proposed model avoids the polysemy problem where spurious associations across terms in different contexts are established due to the presence of one or more common polysemic terms. The algorithm works in an incremental fashion where energy values are adjusted after each document is added to the corpus. This has the advantage that computation is localized around the set of terms contained in the new document, thus making the algorithm run much faster than conventional matrix computations used for singular value decompositions.